Blame for README.md

---
tags:
- multabench
- embeddings
---

# MulTaBench text embeddings (TILE)

Part of [MulTaBench](https://github.com/alanarazi7/MulTaBench), a benchmark for multimodal tabular learning.

Frozen [TILE](https://github.com/soda-inria/tile) embeddings of the MulTaBench text datasets (Llama-3.1-8B, first 14
layers), cached so the TILE models' leaderboard runs don't re-encode them. TILE reads whole rows, so each
`<dataset>/embeddings.npz` maps every distinct serialized row to one mean-pooled 4096-d vector per string column, the
text columns and the ones stored as categories (float16; PCA is fitted per run, on the training rows); `meta.json`
holds the dataset's rows and columns, the encoding time per column and the hardware. See
[Cached embeddings](https://github.com/alanarazi7/MulTaBench#cached-embeddings). An earlier cache, of the text columns
only, is kept in [multabench/embeddings-tile-legacy](https://huggingface.co/datasets/multabench/embeddings-tile-legacy).

**Paper:** [MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image](https://huggingface.co/papers/2605.10616) · **Leaderboard:** [multabench/leaderboard](https://huggingface.co/spaces/multabench/leaderboard)

## Citation

```bibtex
@misc{arazi2026multabenchbenchmarkingmultimodaltabular,
  title={MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image},
  author={Alan Arazi and Eilam Shapira and Shoham Grunblat and Mor Ventura and Elad Hoffer and Gioia Blayer and David Holzmüller and Lennart Purucker and Gaël Varoquaux and Frank Hutter and Roi Reichart},
  year={2026},
  eprint={2605.10616},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2605.10616},
}
```